Today, immensely large amounts of visual data are captured and recorded every moment. Also, we can download almost an infinite number of images and videos from the Internet, where the stored visual data are explosively increasing.

 

[ Can the large amount of visual data produce something new?]

According to the idea of “quantity breeds quality”, researchers are trying to find out new phenomena and their applications on the large amounts of visual data. “Generic object recognition”, “hallucination”, “irregularity detection”, and “cascaded ADA Boosting” are the successful examples. We already have these applications, however, they can be extended further.

 

[[Arm ourselves for fighting with the monster!]

Through these researches, we noticed that most algorithms performing search, clustering, regression, and classification proposed so far lose effect on a large amount of visual data. This implies that more scalable algorithms and architectures have to be developed. For example, compact and powerful feature descriptors, memory efficient algorithms, distributed parallel architectures, and so on.

 

[We need metallurgy for making strong armor.]

 For utilizing the visual data, label data are necessary in many cases. However, labels are always expensive and available labels are often noisy. How can we make robust algorithms against such noisy labels?

Also, an imbalanced training set may introduce bias into classifiers. How can we remove it?

 

WE WANT TO HAVE A WORKSHOP FOR SOLVING THESE PROBLEMS!

 

Paper submission deadline	June 15, 2009 June 23, 2009
Paper acceptance notification	July 25, 2009
Workshop	October 4, 2009